Multi-Class Segmentation with Relative Location Prior

摘要: Multi-class image segmentation has made significant advances in recent years through the combination of local and global features. One important type feature is that inter-class spatial relationships. For example, identifying "tree" pixels indicates above to sides are more likely be "sky" whereas below "grass." Incorporating such information across entire between all classes a computational challenge as it image-dependent, hence, cannot precomputed. In this work we propose method for capturing from relationships encoding feature. We employ two-stage classification process label pixels. First, generate predictions which used compute relative location learned maps. In second stage, combine with appearance-based features provide final segmentation. compare our results published on several multi-class databases show incorporation allows us significantly outperform current state-of-the-art.